In depth
Multi-agent setups are how production-grade agent work scales. The canonical pattern is a coordinator that decomposes a task and delegates each piece to a specialist worker, then merges the results. Each agent has its own brain, its own tools, its own system prompt, which makes it easy to route cheap turns to a cheap model and premium turns to a premium one. The cost story alone is usually enough to justify going multi-agent on any non-trivial task.
What it looks like in practice
YAML
1agents:2 - id: coordinator3 brain: { model: claude-sonnet-4-6 }4 workers: [explorer]5 - id: explorer6 brain: { model: claude-haiku-4-5 }7 modules: [filesystem]Related concepts
Coordinator agentThe top-level agent in a multi-agent setup. Plans, dispatches workers, merges their results.Specialist agentAn agent with a narrow, focused role spawned by a coordinator.Model routingSending different turn types to different models (cheap for exploration, premium for writing).Agent spawnThe runtime mechanism that lets one agent launch another, optionally in parallel, and wait for the result.
Read the deep dive
How to build a Claude Code clone in YAML
Read article
Newsletter
Get the next post in your inbox.
Engineering notes from the Digitorn team. No marketing, no launch announcements, no "10 prompts that will change your life". Just the things we write that we'd want to read.